What we check before we automate a workflow
Not every workflow should be handed to AI, and the ones that should rarely look the way people expect. Here is the short checklist we run first.
Enthusiasm for AI tends to point at the wrong workflows first. The flashy demo target and the high-value real target are usually not the same thing. Before we put AI in the path of any workflow, we run a short set of checks. They are unglamorous, and they save months.
Is the work mostly reading and reasoning over text?
AI earns its keep where the bottleneck is comprehension: reading a codebase, correlating logs, summarizing a thread, drafting from a spec. If the real constraint is something else, a slow external system, an approval that has to be human, a decision that is political rather than technical, then automating the text part buys you very little. Find the actual constraint first.
Can a wrong answer be caught before it does damage?
A fast wrong answer is worse than a slow right one. Before we automate, we ask where a confident mistake would go and whether anything would catch it. If the workflow has a natural checkpoint, a test suite, a review, a validation query, a human sign-off, it is a strong candidate. If a wrong output flows straight to a customer or a ledger with nothing in between, we build the checkpoint before we build the automation.
Does the workflow happen often enough to matter?
A workflow you run twice a year is rarely worth the investment to automate, no matter how annoying it is each time. The compounding value lives in the frequent, boring, every-day workflows. Those are also the ones people forget to mention, because they have stopped noticing the time they cost.
Will the team actually adopt it?
The best workflow in the world is worthless if it sits beside how people really work instead of inside it. We check whether the change fits existing tools and habits, or whether we are asking people to detour. If it is a detour, adoption quietly fails and the old way wins. Real adoption usually means meeting the team where they already are.
Can we measure whether it worked?
If we cannot say what “better” looks like in numbers before we start, we are guessing. We agree on the measure up front: time to merge, time to answer, incidents caught earlier, hours returned. Then we can tell the difference between a change that felt good and a change that was good.
None of this is exotic. It is the discipline of pointing a powerful tool at the right target, which is most of what separates an AI initiative that pays off from one that produces a lot of impressive activity and no movement.
By Activated
Want this applied to your delivery, not just discussed?
Book a working session